model_selection

Perform a recursion of admissible model subsets for a given model ensemble

Specification

  • Alias: None

  • Arguments: None

Description

For the ACV ([GGEJ20]) and generalized ACV ([BLWL22]) methods, this option specifies an enumerative search over all model subsets for a given model ensemble. In the ACV case, each model subset employs a single DAG definition (a “peer” DAG where each active approximation node points to the root node); thus this keyword is only used to search over possible subsets given this fixed ACV DAG. For generalized ACV, model selection may be combined with enumeration of admissible DAGs using different DAG recursion throttles.

The model subset (and DAG definition) with the best performance (lowest estimator variance for a prescribed budget or lowest cost for a prescribed accuracy) is selected for final post-processing.

Examples

method,
    model_pointer = 'ENSEMBLE'
        approximate_control_variate acv_mf
      pilot_samples = 50 seed = 8674132
      search_model_graphs
        no_recursion                      # ACV case
#       kl_recursion                      # GenACV case 1 of 3
#       partial_recursion depth_limit = 2 # GenACV case 2 of 3
#       full_recursion                    # GenACV case 3 of 3
        model_selection                   # this option
      max_function_evaluations = 500